The biased co-counsel: analyzing gendered defense strategies in AI-assisted criminal litigation

Abstract As artificial intelligence (AI) assumes increasingly complex advisory roles in legal practice, its capacity to navigate the nuances of criminal defense remains a critical yet underexplored area. This study investigates the strategic and ethical limitations of large language models (LLMs) by deploying ChatGPT-5 as a simulated defense attorney in a sexual assault scenario governed by the Greek Penal Code. Using a controlled experimental design with 90 iterations, the research systematically varied the gender identity of the accuser (woman, man, and transgender woman) to examine differences in the generated defense narratives. The results indicate significant tensions between probabilistic text generation and Continental legal frameworks. Doctrinally, the model frequently bypassed contemporary consent-based statutes in favor of outdated behavioral indicators and exhibited limitations in accommodating essential procedural safeguards, such as the active participation of the civil claimant. Furthermore, the system demonstrated a consistent pattern of constructing defense motives based on demographic variables, attributing relational vulnerability to women, calculated opportunism to men, and ideological agendas to trans women, while defaulting to a male defendant. These findings illustrate that although AI can generate persuasive legal rhetoric, it inherently embeds cultural and societal biases within its outputs. The study underscores the critical importance of human oversight and procedural safeguards, highlighting that the pursuit of efficiency in legal technology must not compromise doctrinal integrity or fairness in legal decision-making.